[SPARK-39180][SQL] Simplify the planning of limit and offset

cloud-fan · cloud-fan · commit d9129205a013 · 2022-05-17T01:09:27.000+08:00
### What changes were proposed in this pull request? This PR simplifies the planning of limit and offset: 1. Unify the semantics of physical plans that need to deal with limit + offset. These physical plans always do limit first, then offset. The planner rule should set limit and offset properly, for different plans, such as limit + offset and offset + limit. 2. Refactor the planner rule `SpecialLimit` to reuse the code of planning `TakeOrderedAndProjectExec`. 3. Let `GlobalLimitExec` to handle offset as well, so that we can remove `GlobalLimitAndOffsetExec`. This matches `CollectLimitExec`. ### Why are the changes needed? code simplification ### Does this PR introduce _any_ user-facing change? no ### How was this patch tested? existing tests Closes #36541 from cloud-fan/offset. Lead-authored-by: Wenchen Fan <cloud0fan@gmail.com> Co-authored-by: Wenchen Fan <wenchen@databricks.com> Signed-off-by: Wenchen Fan <wenchen@databricks.com>
diff --git a/sql/catalyst/src/main/scala/org/apache/spark/sql/catalyst/plans/logical/basicLogicalOperators.scala b/sql/catalyst/src/main/scala/org/apache/spark/sql/catalyst/plans/logical/basicLogicalOperators.scala
@@ -1303,12 +1303,25 @@ case class LocalLimit(limitExpr: Expression, child: LogicalPlan) extends OrderPr
     copy(child = newChild)
 }
 
+object OffsetAndLimit {
+  def unapply(p: GlobalLimit): Option[(Int, Int, LogicalPlan)] = {
+    p match {
+      // Optimizer pushes local limit through offset, so we need to match the plan this way.
+      case GlobalLimit(IntegerLiteral(globalLimit),
+             Offset(IntegerLiteral(offset),
+               LocalLimit(IntegerLiteral(localLimit), child)))
+          if globalLimit + offset == localLimit =>
+        Some((offset, globalLimit, child))
+      case _ => None
+    }
+  }
+}
+
 object LimitAndOffset {
-  def unapply(p: GlobalLimit): Option[(Expression, Expression, LogicalPlan)] = {
+  def unapply(p: Offset): Option[(Int, Int, LogicalPlan)] = {
     p match {
-      case GlobalLimit(le1, Offset(le2, LocalLimit(le3, child))) if le1.eval().asInstanceOf[Int]
-        + le2.eval().asInstanceOf[Int] == le3.eval().asInstanceOf[Int] =>
-        Some((le1, le2, child))
+      case Offset(IntegerLiteral(offset), Limit(IntegerLiteral(limit), child)) =>
+        Some((limit, offset, child))
       case _ => None
     }
   }
diff --git a/sql/core/src/main/scala/org/apache/spark/sql/execution/SparkStrategies.scala b/sql/core/src/main/scala/org/apache/spark/sql/execution/SparkStrategies.scala
@@ -81,55 +81,56 @@ abstract class SparkStrategies extends QueryPlanner[SparkPlan] {
    */
   object SpecialLimits extends Strategy {
     override def apply(plan: LogicalPlan): Seq[SparkPlan] = plan match {
-      case ReturnAnswer(rootPlan) => rootPlan match {
-        case Limit(IntegerLiteral(limit), Sort(order, true, child))
-            if limit < conf.topKSortFallbackThreshold =>
-          TakeOrderedAndProjectExec(limit, order, child.output, planLater(child)) :: Nil
-        case Limit(IntegerLiteral(limit), Project(projectList, Sort(order, true, child)))
-            if limit < conf.topKSortFallbackThreshold =>
-          TakeOrderedAndProjectExec(limit, order, projectList, planLater(child)) :: Nil
+      // Call `planTakeOrdered` first which matches a larger plan.
+      case ReturnAnswer(rootPlan) => planTakeOrdered(rootPlan).getOrElse(rootPlan match {
+        // We should match the combination of limit and offset first, to get the optimal physical
+        // plan, instead of planning limit and offset separately.
+        case LimitAndOffset(limit, offset, child) =>
+          CollectLimitExec(limit = limit, child = planLater(child), offset = offset)
+        case OffsetAndLimit(offset, limit, child) =>
+          // 'Offset a' then 'Limit b' is the same as 'Limit a + b' then 'Offset a'.
+          CollectLimitExec(limit = offset + limit, child = planLater(child), offset = offset)
         case Limit(IntegerLiteral(limit), child) =>
-          CollectLimitExec(limit, planLater(child)) :: Nil
-        case LimitAndOffset(IntegerLiteral(limit), IntegerLiteral(offset),
-          Sort(order, true, child)) if limit + offset < conf.topKSortFallbackThreshold =>
-          TakeOrderedAndProjectExec(
-            limit, order, child.output, planLater(child), offset) :: Nil
-        case LimitAndOffset(IntegerLiteral(limit), IntegerLiteral(offset),
-          Project(projectList, Sort(order, true, child)))
-            if limit + offset < conf.topKSortFallbackThreshold =>
-          TakeOrderedAndProjectExec(
-            limit, order, projectList, planLater(child), offset) :: Nil
-        case LimitAndOffset(IntegerLiteral(limit), IntegerLiteral(offset), child) =>
-          CollectLimitExec(limit, planLater(child), offset) :: Nil
+          CollectLimitExec(limit = limit, child = planLater(child))
         case logical.Offset(IntegerLiteral(offset), child) =>
-          CollectLimitExec(child = planLater(child), offset = offset) :: Nil
+          CollectLimitExec(child = planLater(child), offset = offset)
         case Tail(IntegerLiteral(limit), child) =>
-          CollectTailExec(limit, planLater(child)) :: Nil
-        case other => planLater(other) :: Nil
-      }
+          CollectTailExec(limit, planLater(child))
+        case other => planLater(other)
+      })  :: Nil
+
+      case other => planTakeOrdered(other).toSeq
+    }
+
+    private def planTakeOrdered(plan: LogicalPlan): Option[SparkPlan] = plan match {
+      // We should match the combination of limit and offset first, to get the optimal physical
+      // plan, instead of planning limit and offset separately.
+      case LimitAndOffset(limit, offset, Sort(order, true, child))
+          if limit < conf.topKSortFallbackThreshold =>
+        Some(TakeOrderedAndProjectExec(
+          limit, order, child.output, planLater(child), offset))
+      case LimitAndOffset(limit, offset, Project(projectList, Sort(order, true, child)))
+          if limit < conf.topKSortFallbackThreshold =>
+        Some(TakeOrderedAndProjectExec(
+          limit, order, projectList, planLater(child), offset))
+      // 'Offset a' then 'Limit b' is the same as 'Limit a + b' then 'Offset a'.
+      case OffsetAndLimit(offset, limit, Sort(order, true, child))
+          if offset + limit < conf.topKSortFallbackThreshold =>
+        Some(TakeOrderedAndProjectExec(
+          offset + limit, order, child.output, planLater(child), offset))
+      case OffsetAndLimit(offset, limit, Project(projectList, Sort(order, true, child)))
+          if offset + limit < conf.topKSortFallbackThreshold =>
+        Some(TakeOrderedAndProjectExec(
+          offset + limit, order, projectList, planLater(child), offset))
       case Limit(IntegerLiteral(limit), Sort(order, true, child))
           if limit < conf.topKSortFallbackThreshold =>
-        TakeOrderedAndProjectExec(limit, order, child.output, planLater(child)) :: Nil
+        Some(TakeOrderedAndProjectExec(
+          limit, order, child.output, planLater(child)))
       case Limit(IntegerLiteral(limit), Project(projectList, Sort(order, true, child)))
           if limit < conf.topKSortFallbackThreshold =>
-        TakeOrderedAndProjectExec(limit, order, projectList, planLater(child)) :: Nil
-      // This is a global LIMIT and OFFSET over a logical sorting operator,
-      // where the sum of specified limit and specified offset is less than a heuristic threshold.
-      // In this case we generate a physical top-K sorting operator, passing down
-      // the limit and offset values to be evaluated inline during the physical
-      // sorting operation for greater efficiency.
-      case LimitAndOffset(IntegerLiteral(limit), IntegerLiteral(offset),
-        Sort(order, true, child)) if limit + offset < conf.topKSortFallbackThreshold =>
-          TakeOrderedAndProjectExec(
-            limit, order, child.output, planLater(child), offset) :: Nil
-      case LimitAndOffset(IntegerLiteral(limit), IntegerLiteral(offset),
-        Project(projectList, Sort(order, true, child)))
-          if limit + offset < conf.topKSortFallbackThreshold =>
-        TakeOrderedAndProjectExec(limit, order, projectList, planLater(child), offset) :: Nil
-      case LimitAndOffset(IntegerLiteral(limit), IntegerLiteral(offset), child) =>
-        GlobalLimitAndOffsetExec(limit, offset, planLater(child)) :: Nil
-      case _ =>
-        Nil
+        Some(TakeOrderedAndProjectExec(
+          limit, order, projectList, planLater(child)))
+      case _ => None
     }
   }
 
@@ -814,12 +815,19 @@ abstract class SparkStrategies extends QueryPlanner[SparkPlan] {
       case logical.LocalRelation(output, data, _) =>
         LocalTableScanExec(output, data) :: Nil
       case CommandResult(output, _, plan, data) => CommandResultExec(output, plan, data) :: Nil
+      // We should match the combination of limit and offset first, to get the optimal physical
+      // plan, instead of planning limit and offset separately.
+      case LimitAndOffset(limit, offset, child) =>
+        GlobalLimitExec(limit, planLater(child), offset) :: Nil
+      case OffsetAndLimit(offset, limit, child) =>
+        // 'Offset a' then 'Limit b' is the same as 'Limit a + b' then 'Offset a'.
+        GlobalLimitExec(limit = offset + limit, child = planLater(child), offset = offset) :: Nil
       case logical.LocalLimit(IntegerLiteral(limit), child) =>
         execution.LocalLimitExec(limit, planLater(child)) :: Nil
       case logical.GlobalLimit(IntegerLiteral(limit), child) =>
         execution.GlobalLimitExec(limit, planLater(child)) :: Nil
       case logical.Offset(IntegerLiteral(offset), child) =>
-        GlobalLimitAndOffsetExec(offset = offset, child = planLater(child)) :: Nil
+        GlobalLimitExec(child = planLater(child), offset = offset) :: Nil
       case union: logical.Union =>
         execution.UnionExec(union.children.map(planLater)) :: Nil
       case g @ logical.Generate(generator, _, outer, _, _, child) =>
diff --git a/sql/core/src/main/scala/org/apache/spark/sql/execution/limit.scala b/sql/core/src/main/scala/org/apache/spark/sql/execution/limit.scala
@@ -37,11 +37,11 @@ trait LimitExec extends UnaryExecNode {
 }
 
 /**
- * Take the first `limit` + `offset` elements and collect them to a single partition and then to
- * drop the first `offset` elements.
+ * Take the first `limit` elements, collect them to a single partition and then to drop the
+ * first `offset` elements.
  *
- * This operator will be used when a logical `Limit` operation is the final operator in an
- * logical plan, which happens when the user is collecting results back to the driver.
+ * This operator will be used when a logical `Limit` and/or `Offset` operation is the final operator
+ * in an logical plan, which happens when the user is collecting results back to the driver.
  */
 case class CollectLimitExec(limit: Int = -1, child: SparkPlan, offset: Int = 0) extends LimitExec {
   assert(limit >= 0 || (limit == -1 && offset > 0))
@@ -56,7 +56,7 @@ case class CollectLimitExec(limit: Int = -1, child: SparkPlan, offset: Int = 0)
     // Then [1, 2, 3] will be taken and output [3].
     if (limit >= 0) {
       if (offset > 0) {
-        child.executeTake(limit + offset).drop(offset)
+        child.executeTake(limit).drop(offset)
       } else {
         child.executeTake(limit)
       }
@@ -79,11 +79,7 @@ case class CollectLimitExec(limit: Int = -1, child: SparkPlan, offset: Int = 0)
         childRDD
       } else {
         val locallyLimited = if (limit >= 0) {
-          if (offset > 0) {
-            childRDD.mapPartitionsInternal(_.take(limit + offset))
-          } else {
-            childRDD.mapPartitionsInternal(_.take(limit))
-          }
+          childRDD.mapPartitionsInternal(_.take(limit))
         } else {
           childRDD
         }
@@ -98,7 +94,7 @@ case class CollectLimitExec(limit: Int = -1, child: SparkPlan, offset: Int = 0)
       }
       if (limit >= 0) {
         if (offset > 0) {
-          singlePartitionRDD.mapPartitionsInternal(_.slice(offset, offset + limit))
+          singlePartitionRDD.mapPartitionsInternal(_.slice(offset, limit))
         } else {
           singlePartitionRDD.mapPartitionsInternal(_.take(limit))
         }
@@ -164,8 +160,8 @@ trait BaseLimitExec extends LimitExec with CodegenSupport {
 
   override def outputOrdering: Seq[SortOrder] = child.outputOrdering
 
-  protected override def doExecute(): RDD[InternalRow] = child.execute().mapPartitions { iter =>
-    iter.take(limit)
+  protected override def doExecute(): RDD[InternalRow] = child.execute().mapPartitionsInternal {
+    iter => iter.take(limit)
   }
 
   override def inputRDDs(): Seq[RDD[InternalRow]] = {
@@ -215,61 +211,52 @@ case class LocalLimitExec(limit: Int, child: SparkPlan) extends BaseLimitExec {
 }
 
 /**
- * Take the first `limit` elements of the child's single output partition.
+ * Take the first `limit` elements and then drop the first `offset` elements in the child's single
+ * output partition.
  */
-case class GlobalLimitExec(limit: Int, child: SparkPlan) extends BaseLimitExec {
-
-  override def requiredChildDistribution: List[Distribution] = AllTuples :: Nil
-
-  override protected def withNewChildInternal(newChild: SparkPlan): SparkPlan =
-    copy(child = newChild)
-}
-
-/**
- * Skip the first `offset` elements then take the first `limit` of the following elements in
- * the child's single output partition.
- */
-case class GlobalLimitAndOffsetExec(
-    limit: Int = -1,
-    offset: Int,
-    child: SparkPlan) extends BaseLimitExec {
-  assert(offset > 0)
+case class GlobalLimitExec(limit: Int = -1, child: SparkPlan, offset: Int = 0)
+  extends BaseLimitExec {
+  assert(limit >= 0 || (limit == -1 && offset > 0))
 
   override def requiredChildDistribution: List[Distribution] = AllTuples :: Nil
 
-  override def doExecute(): RDD[InternalRow] = if (limit >= 0) {
-    child.execute().mapPartitionsInternal(iter => iter.slice(offset, limit + offset))
-  } else {
-    child.execute().mapPartitionsInternal(iter => iter.drop(offset))
+  override def doExecute(): RDD[InternalRow] = {
+    if (offset > 0) {
+      if (limit >= 0) {
+        child.execute().mapPartitionsInternal(iter => iter.slice(offset, limit))
+      } else {
+        child.execute().mapPartitionsInternal(iter => iter.drop(offset))
+      }
+    } else {
+      super.doExecute()
+    }
   }
 
-  private lazy val skipTerm = BaseLimitExec.newLimitCountTerm()
-
   override def doConsume(ctx: CodegenContext, input: Seq[ExprCode], row: ExprCode): String = {
-    ctx.addMutableState(
-      CodeGenerator.JAVA_INT, skipTerm, forceInline = true, useFreshName = false)
-    if (limit >= 0) {
-      // The counter name is already obtained by the upstream operators via `limitNotReachedChecks`.
-      // Here we have to inline it to not change its name. This is fine as we won't have many limit
-      // operators in one query.
-      ctx.addMutableState(
-        CodeGenerator.JAVA_INT, countTerm, forceInline = true, useFreshName = false)
-      s"""
-         | if ($skipTerm < $offset) {
-         |   $skipTerm += 1;
-         | } else if ($countTerm < $limit) {
-         |   $countTerm += 1;
-         |   ${consume(ctx, input)}
-         | }
+    if (offset > 0) {
+      val skipTerm = ctx.addMutableState(CodeGenerator.JAVA_INT, "rowsSkipped", forceInline = true)
+      if (limit > 0) {
+        // In codegen, we skip the first `offset` rows, then take the first `limit - offset` rows.
+        val finalLimit = limit - offset
+        s"""
+           | if ($skipTerm < $offset) {
+           |   $skipTerm += 1;
+           | } else if ($countTerm < $finalLimit) {
+           |   $countTerm += 1;
+           |   ${consume(ctx, input)}
+           | }
          """.stripMargin
+      } else {
+        s"""
+           | if ($skipTerm < $offset) {
+           |   $skipTerm += 1;
+           | } else {
+           |   ${consume(ctx, input)}
+           | }
+         """.stripMargin
+      }
     } else {
-      s"""
-         | if ($skipTerm < $offset) {
-         |   $skipTerm += 1;
-         | } else {
-         |   ${consume(ctx, input)}
-         | }
-       """.stripMargin
+      super.doConsume(ctx, input, row)
     }
   }
 
@@ -278,9 +265,9 @@ case class GlobalLimitAndOffsetExec(
 }
 
 /**
- * Take the first limit elements as defined by the sortOrder, and do projection if needed.
- * This is logically equivalent to having a Limit operator after a [[SortExec]] operator,
- * or having a [[ProjectExec]] operator between them.
+ * Take the first `limit` elements as defined by the sortOrder, then drop the first `offset`
+ * elements, and do projection if needed. This is logically equivalent to having a Limit and/or
+ * Offset operator after a [[SortExec]] operator, or having a [[ProjectExec]] operator between them.
  * This could have been named TopK, but Spark's top operator does the opposite in ordering
  * so we name it TakeOrdered to avoid confusion.
  */
@@ -297,12 +284,8 @@ case class TakeOrderedAndProjectExec(
 
   override def executeCollect(): Array[InternalRow] = {
     val ord = new LazilyGeneratedOrdering(sortOrder, child.output)
-    val data = if (offset > 0) {
-      child.execute().mapPartitionsInternal(_.map(_.copy()))
-        .takeOrdered(limit + offset)(ord).drop(offset)
-    } else {
-      child.execute().mapPartitionsInternal(_.map(_.copy())).takeOrdered(limit)(ord)
-    }
+    val limited = child.execute().mapPartitionsInternal(_.map(_.copy())).takeOrdered(limit)(ord)
+    val data = if (offset > 0) limited.drop(offset) else limited
     if (projectList != child.output) {
       val proj = UnsafeProjection.create(projectList, child.output)
       data.map(r => proj(r).copy())
@@ -328,15 +311,10 @@ case class TakeOrderedAndProjectExec(
       val singlePartitionRDD = if (childRDD.getNumPartitions == 1) {
         childRDD
       } else {
-        val localTopK = if (offset > 0) {
-          childRDD.mapPartitionsInternal { iter =>
-            Utils.takeOrdered(iter.map(_.copy()), limit + offset)(ord)
-          }
-        } else {
-          childRDD.mapPartitionsInternal { iter =>
-            Utils.takeOrdered(iter.map(_.copy()), limit)(ord)
-          }
+        val localTopK = childRDD.mapPartitionsInternal { iter =>
+          Utils.takeOrdered(iter.map(_.copy()), limit)(ord)
         }
+
         new ShuffledRowRDD(
           ShuffleExchangeExec.prepareShuffleDependency(
             localTopK,
@@ -347,11 +325,8 @@ case class TakeOrderedAndProjectExec(
           readMetrics)
       }
       singlePartitionRDD.mapPartitionsInternal { iter =>
-        val topK = if (offset > 0) {
-          Utils.takeOrdered(iter.map(_.copy()), limit + offset)(ord).drop(offset)
-        } else {
-          Utils.takeOrdered(iter.map(_.copy()), limit)(ord)
-        }
+        val limited = Utils.takeOrdered(iter.map(_.copy()), limit)(ord)
+        val topK = if (offset > 0) limited.drop(offset) else limited
         if (projectList != child.output) {
           val proj = UnsafeProjection.create(projectList, child.output)
           topK.map(r => proj(r))
